PV Module Classification Based on Spatial Distribution of Soiling Using Deep Learning Model

Soiling is the most common phenomenon, especially in desert regions, resulting in power losses and reduced reliability for PV modules. The spatial distribution of dust particles can be uniform or non-uniform, both negatively affecting the performance. However, the latter has more pronounced effects as it leads to partial shading, leading to current mismatching with reduced performance and the formation of hot spots. Hence, it is necessary to perform timely inspection of solar PV farms to detect and identify the distribution of dust. We propose a deep learning model for detecting dust levels on a PV module based on uniformity to achieve this objective. The workflow involves first segmenting the modules from the farm, followed by its individual localization, classification based on dust or no dust, and finally, classification based on uniformity of the dust deposition. The data is gathered using a drone equipped with a camera autonomously navigated using real-time kinetics (RTK). The overall accuracy of the classification model is more than 90 %. This enables a faster inspection process, which, combined with our algorithm, improves the overall efficiency and accuracy of the maintenance procedure.